How to judge quantum advantage claims today

If you want to know whether quantum advantage claims are real, stop reading press releases and look at how fast classical methods answer the same problem. In several documented cases, a classical or AI-based method caught up with a quantum result quickly, and in one case a Commodore 64 replicated a result published from an IBM quantum device.

Where quantum advantage stands in 2026

Quantum advantage in 2026 remains an argument about evidence rather than a settled milestone. No fault-tolerant quantum computer exists, and current devices are not economically useful because classical methods still compete on the problems enterprises care about. The debate is now about how fast classical algorithms overtake each quantum demonstration.

The term itself is unstable. Jay Gambetta, who leads IBM Quantum, and Hartmut Neven, who leads Google Quantum AI, both spoke at an MIT panel on the evolution of quantum and disagreed productively about what counts. Gambetta said plainly that none of the entries on the field's public advantage trackers show advantage. Neven argued that devices doing science that classical machines cannot easily simulate is already a legitimate result, separate from return on investment.

That split matters for how you read any claim. A scientific advantage, where a device produces a result hard to reproduce classically, is a weaker and different claim than an economic advantage, where a business would rationally pay to use the machine. Most 2026 headlines blur the two.

Both leaders also agreed on one boundary: the fault-tolerant machine they are building does not exist yet. Everything below should be read against that admission.

Why rational exuberance beats the hype label

Rational exuberance, a phrase Neven attributes to economist Mohamed El-Erian, describes an investment cycle where the technology is real and some investors still lose money by arriving late. Railroads and the early internet followed that pattern, which is different from a fad that leaves nothing behind.

The practical consequence is that hype and substance can coexist in the same market. AI is now used daily by a large share of knowledge workers, which supports the claim that the underlying technology is not empty. Quantum has not reached that point, and the panel did not pretend otherwise.

Gambetta's framing was harsher on current hardware. He compared today's quantum computers to early 1980s home machines such as the Sinclair ZX81 and BBC Micro: historically important, widely owned, and not the machines that later ran the economy.

The useful question is therefore not whether quantum is overhyped. It is which specific claims are falsifiable now, and which depend on hardware that has not been built.

The classical rebuttals that changed the story

Repeated classical rebuttals have shaped how the field presents results, and several of them arrived within weeks or months of the original quantum paper. That back-and-forth is now part of the evidence record rather than an embarrassment.

Three episodes were discussed on the panel, each with a different lesson.

The Commodore 64 replication

An IBM paper reported results on fundamental physics questions about the Ising model using a quantum device. A student later reproduced the same result on a Commodore 64, a 1980s home computer, without quantum hardware. IBM continued the research program and iterated rather than abandoning it, and the company's advantage tracker now reports results under its own name. Gambetta's own assessment at the panel was that none of the tracker entries demonstrate advantage.

The peak circuit race

Under what Scott Aaronson, a theoretical computer scientist at the University of Texas at Austin, has called the peak circuit framing, a hidden signal is planted in a circuit and the task is to locate it. A quantum device found the peak faster than a tensor-network method. An AI-based black-box method then found it roughly 100 times faster. Quantinuum and BlueQubit have since posted results finding the peak faster on quantum hardware.

Observable estimation

Gambetta's team estimated observables from circuits resembling correlated dynamics, and reported observables that flatiron tensor methods could not simulate. A student then used a single-iteration approach that removed much of the interference and matched the result classically.

The pattern is consistent enough to be a rule: a quantum demonstration is a snapshot, and classical methods move fast.

The 15-spin to 50-spin gap in chemistry

Quantum chemistry is where the nearest practical use case sits, and the gap between current demonstrations and useful work is measurable. The panel put the crossover at roughly 50 spins, the point at which classical simulation becomes very expensive.

Google's team demonstrated the approach on small molecules of about 15 spins, learning properties that were unknown before the experiment. On those examples a classical machine could still have done the work, so the demonstration shows the method rather than a practical win.

Neven's own estimate was that an NMR practitioner who does not care about quantum computing at all might choose a quantum device as the best tool within about two years, provided error rates fall enough to add computational volume. He also conceded that he is not an NMR specialist and cannot promise the community will adopt the protocol.

The important point for a reader evaluating claims: the molecule size, the spin count and the error rate are the variables that matter. A headline about quantum chemistry is not evidence without them.

The nearest practical quantum chemistry use case is a molecule of more than 50 spins, versus roughly 15 spins demonstrated. The distance between those numbers is what error-rate improvements have to close.

Hardware bets, error correction and the CMOS question

Superconducting qubits are the leading industrial bet, chosen because the surrounding microwave, radar and CMOS industries already exist. That choice shapes the cost estimate: reusing existing manufacturing infrastructure is cheaper than rebuilding a semiconductor industry from scratch.

The architecture is not settled. Google has invested in QuEra, a neutral-atom quantum computing company, partly to keep a close view of a competing modality. Recent work has also moved beyond aluminum, the standard metal in superconducting circuits, toward tantalum, which has reported better coherence in published experiments.

Error correction is where the theory-to-practice transfer has been fastest. The surface code, the dominant error-correcting scheme for years, is now being challenged by newer code families that need fewer physical qubits for a given logical error rate. Both leaders pointed to this as the change most likely to produce more effective machines.

The bottlenecks are also physical. Gambetta described a system with 10 cryogenic fridges connected together, where a single failure forces a cooldown of the whole set. His proposed fix is a vacuum chamber that can be opened and closed at 4 kelvin, a temperature regime where space-physics engineering has already shown the approach is possible, even though the same operation at millikelvin remains impractical.

A comparison of the main claims helps separate what is measured from what is projected.

Claim	Status	Evidence
Fault-tolerant quantum computer	Not built	Neither leader claims one exists
Quantum chemistry on 15-spin molecules	Demonstrated	Classical simulation still possible at that size
Quantum chemistry past 50 spins	Projected, roughly two years	Depends on lower error rates
Quantum advantage on public trackers	Disputed	Google and IBM leaders disagree on interpretation
Surface code as the standard	Being challenged	Newer codes claim fewer physical qubits

What students and researchers can work on

The panel's advice to researchers was specific: pick problems before they become crowded, and work on the parts of the stack that well-funded labs cannot cover alone. Both speakers treated quantum computing as a full-stack systems engineering problem with too many trade-offs for any single organization to explore.

Concrete directions that came up:

1. Quantum information science, to extract more from existing devices through error correction and simpler algorithms.
2. Hybrid algorithms that treat the quantum processor as a subroutine inside classical high-performance computing, which requires the two communities to work together.
3. Peripheral engineering, especially cryogenic interfaces that let a system be opened and serviced without warming the whole machine.
4. Materials and fabrication experiments that large labs cannot justify, such as alternative superconducting metals.
5. New algorithms beyond the established ones for factoring and simulation, which both speakers said remain under-explored.

Gambetta's career path is the supporting anecdote for the first point. He moved from laser spectroscopy into theory, then into superconducting qubits, changing subfield each time he found something he did not understand. His stated rule was to pick the hardest problems available and find mentors who will take a chance on an outsider.

Neven's version of the same advice was timing-based: enter areas at the edge of current interest rather than after they become crowded. He worked on what would now be called self-driving cars in the 1990s and on AI in the early 2000s.

The cost question behind all of this is unsettled. Neven disputed the premise that reaching a fault-tolerant machine requires trillions of dollars and put the figure at orders of billions. Gambetta argued the total depends on whether the field reinvents the semiconductor industry or reuses it, and said the first quantum computers may mainly teach the field how to build later ones.

FAQ

• Does a fault-tolerant quantum computer exist in 2026? No. As of September 2026 no fault-tolerant quantum computer has been built, and both IBM Quantum and Google Quantum AI describe it as the goal their current roadmaps target. Current devices run with error rates high enough that error correction is required to do useful work.

• Is quantum advantage real today? Not in an economic sense. Current devices are not competitive with classical computing for commercial problems, and a classical or AI-based method has matched several published quantum demonstrations. Scientific advantage, meaning a result that is hard to reproduce classically, is a separate and more defensible claim.

• How many spins does a molecule need before quantum computing helps? Roughly 50 spins is the point where classical simulation becomes very expensive, according to the estimate given at the MIT panel. Quantum chemistry demonstrations have reached about 15 spins, where classical machines can still do the job.

• Which qubit technology is winning? Superconducting qubits are the leading industrial bet because they reuse existing microwave and CMOS manufacturing. Neutral atoms, pursued by companies such as QuEra, are a credible alternative, and major labs invest in them to monitor the competition rather than to replace their own platform.

• What should a student work on in quantum computing? Error correction, hybrid quantum-classical algorithms, cryogenic engineering and algorithm discovery were the areas named on the panel. Both speakers said the field still has far more open questions than solved ones, despite its popularity.

Source video